Source: Mastering Context Engineering — definitive study guide
Prompt engineering gets all the attention. Context engineering is more important. The distinction: prompt engineering is about how you phrase individual instructions. Context engineering is about how you design the entire information environment in which the model operates — what it knows, what it can access, and what it’s currently trying to do.
The three-dimensional context framework: Instructional context is what the model is told to do and how — system prompts, task descriptions, few-shot examples. Informational context is what the model needs to know — retrieved documents (RAG), short and long-term memory, state tracking. Actionable context is what the model can do — tool definitions, tool call results, execution tracking.
Most teams treat context as just the chat log — the sequence of messages sent and received. This misses the Informational and Actionable dimensions entirely. A model without access to relevant retrieved documents will hallucinate the information it needs. A model without memory of previous decisions will make contradictory ones. A model without clean tool definitions will call tools incorrectly. None of these are model intelligence failures. They’re context design failures.
The practical takeaway: when an AI agent fails on a complex task, the first diagnostic question should be “what information was it missing?” not “is the model smart enough?” In most cases, the model was more than capable — it was operating in an information environment that made success nearly impossible. Fix the context first.